DOCSLIB.ORG

Comparative Performance of Machine Learning Algorithms in Cyberbullying Detection: Using Turkish Language Preprocessing Techniques

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Comparative Performance of Machine Learning Algorithms in Cyberbullying Detection: Using Turkish Language Preprocessing Techniques Comparative Performance of Machine Learning Algorithms in Cyberbullying Detection: Using Turkish Language Preprocessing Techniques Emre Cihan ATESa*, Erkan BOSTANCIb, Mehmet Serdar GÜZELc a Department of Security Science, Gendarmerie and Coast Guard Academy (JSGA), Ankara, Turkey. b Department of Computer Engineering, Ankara University, Ankara, Turkey. c Department of Computer Engineering, Ankara University, Ankara, Turkey. Abstract With the increasing use of the internet and social media, it is obvious that cyberbullying has become a major problem. The most basic way for protection against the dangerous consequences of cyberbullying is to actively detect and control the contents containing cyberbullying. When we look at today's internet and social media statistics, it is impossible to detect cyberbullying contents only by human power. Effective cyberbullying detection methods are necessary in order to make social media a safe communication space. Current research efforts focus on using machine learning for detecting and eliminating cyberbullying. Although most of the studies have been conducted on English texts for the detection of cyberbullying, there are few studies in Turkish. Limited methods and algorithms were also used in studies conducted on the Turkish language. In addition, the scope and performance of the algorithms used to classify the texts containing cyberbullying is different, and this reveals the importance of using an appropriate algorithm. The aim of this study is to compare the performance of different machine learning algorithms in detecting Turkish messages containing cyberbullying. In this study, nineteen different classification algorithms were used to identify texts containing cyberbullying using Turkish natural language processing techniques. Precision, recall, accuracy and F1 score values were used to evaluate the performance of classifiers. It was determined that the Light Gradient Boosting Model (LGBM) algorithm showed the best performance with 90.788% accuracy and 90.949% F1 Score value. Keywords — Cyberbullying, machine learning, social media, crime, natural language processing 1. Introduction With the rapid development of social networks today, the scope of internet use has significantly expanded. We are in a period in which people create a virtual personality on social media for purposes such as expressing their emotions and sharing things with other people (Mulyadi and Fitriana, 2018). *Corresponding author. E-mail addresses: [email protected](E.C.Ates), [email protected](E.Bostanci), [email protected](M.S.Guzel). 1 After the virtual personality formation, which usually starts with membership to social media sites, data types such as text, sound, image and video can be transferred. In addition to data transfer, social media is a communication environment where people can also send messages such as messages, status or comments to each other. The rapid increase in the use of social media over the past years has brought many positive and negative developments (Akdemir and Lawless, 2020; Siddiqui and Singh, 2016). One of the most common negative effects is cyberbullying. Cyberbullying can be defined as the act of intentionally violating another person through technological communication tools (Englander et al., 2017; Peter and Petermann, 2018). Harassment or aggression through e-mails, mobile phones, online games and social media can be considered cyberbullying. Since, most studies focusing on detecting cyberbullying are English-based, the Turkish language based studies are quite insufficient. Turkish language, which is being used in Turkey first and foremost, and also many parts of the world, is a head-final language, and that is one of the main factors that make it difficult to do researches with it (Çakir and Güldamlasioǧlu, 2016; Çöltekin, 2020). For example, particularly in Turkey where Turkish is spoken as a mother tongue, when the rate of social media platforms usage in 2020 on which cyberbullying is made are examined (WeAreSocial, 2020); We see that it's the 6th country on Twitter with its 11.800.000 users, the 5th country on Instagram with its 38,000,000 users, the 10th country on Facebook with its 37,000,000 users, and that it's the 15th country on social media platforms with an average usage time of 2 hours 51 minutes a day. When social media usage statistics in Turkey in particular are examined, it reveals how widespread is the use of social media is in Turkey. Therefore, it has become obligatory to detect cyberbullying in Turkish texts. In this study, it is aimed to compare the performance of different machine learning algorithms in cyberbullying detection by using natural language processing techniques specific to Turkish language. Following the introduction of the study, a background on which the concepts of cyber bullying, machine learning, natural language processing and Turkish language are defined respectively has been prepared. In the related works section, studies in the literature on cyberbullying detection are presented. In the methodology section, information was given about the data used for detecting cyberbullying, pre-processing and machine learning models used. In the result and discussion section, the results obtained after machine learning modelling are shared and compared with other studies in the literature. In the conclusion and future works section, the gains obtained after the study and future work plans to increase cyberbullying detection performance in Turkish texts are presented. 2 2. Background 2.1. Cyberbullying Cyberbullying is the state of aggressive and repetitive negative behaviours using technological devices against another person or group with an imbalance of power (Brailovskaia et al., 2018; Englander et al., 2017). The harm caused by cyberbullying is more serious than physical bullying. Items subject to bullying can be dispersed very quickly on the internet and social media environment. Moreover, parallel to the rapid development of technology, the ability to hide the criminals' identities by different methods in virtual platforms and therefore the difficulty in finding criminals creates a situation that motivates criminals to do such acts. Today, regarding the processing of cyberbullying on social media; some cyberbullying behaviours include sending threatening and vulgar messages, disclosing private personal information, creating online gossip, sending repetitive swearing and insulting messages and humiliating any person or group. Cyberbullying can occur with all genders and at different age groups, and the act of cyberbullying can be related to physical, cultural, racial and even religious prejudices (Ferreira et al., 2016; Muralidharan and La Ferle, 2018). It is known that many people experience physical and mental victimization after cyberbullying behaviour. These behaviours also gradually reduce people's tendency to trust social media (Marzano, 2019). For this reason, detecting cyberbullying on social media is a must more than a need. When the Twitter sample of 2020 is examined, an average of 194,444 tweets were sent in different languages every 1 minute (Lewis and Callahan, 2020). In a 24-hour period sample, the average number of tweets sent reached 279,999,360, and it is almost impossible to physically examine such a large amount of data in different languages. For this reason, today's studies focus primarily on normalization of text with natural language processing techniques and subsequently classification based on machine learning. 2.2. Machine Learning Machine learning is the learning ability of a computer to make decisions on its own, based on the existing data and experiences (Hutter et al., 2019). Decisions to be taken within the scope of machine learning are generally for prediction or classification purposes (Sarkar, 2019). Within the scope of this study, machine learning was used to check whether Twitter posts contain cyberbullying or not. In the context, available data is known as training data. The computer trains the models and classifies test data using learning algorithms. There are different learning methods according to the labelling of the training data (Aggarwal, 2018). What we mean by labelling is the classification of data by human power before machine learning. If the learning model is dependent on tagged data, then this model is supervised learning. Unsupervised learning occurs when training data are not tagged. In unsupervised learning, the 3 machine learns to classify itself according to the similarities and differences between the data. The learning to be applied on the combination of tagged and untagged data is semi-supervised learning. Machine learning algorithms are widely used to detect messages containing cyberbullying in social networks (Rosa et al., 2019). Its usage is in the form of supervised learning in general and in this study. There are different learning methods according to the labelling of the training data, and the algorithms used in this study are explained below. Naive Bayes: There is an assumption in this algorithm that the inputs are independent from each other and of equal importance. Although it seems quite simple in terms of the way it works, successful and fast results are obtained in many examples adapted to real life (Liu et al., 2017; Rogel-Salazar, 2018). The fact that the input data is independent means that each attribute is affected by a certain number of inputs and has no significance among them. This situation is usually almost impossible. In general,
Load more

Recommended publications
  • Arxiv:2006.04059V1 [Cs.LG] 7 Jun 2020
    Soft Gradient Boosting Machine Ji Feng1;2, Yi-Xuan Xu1;3, Yuan Jiang3, Zhi-Hua Zhou3 [email protected], fxuyx, jiangy, [email protected] 1Sinovation Ventures AI Institute 2Baiont Technology 3National Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210093, China Abstract Gradient Boosting Machine has proven to be one successful function approximator and has been widely used in a variety of areas. However, since the training procedure of each base learner has to take the sequential order, it is infeasible to parallelize the training process among base learners for speed-up. In addition, under online or incremental learning settings, GBMs achieved sub-optimal performance due to the fact that the previously trained base learners can not adapt with the environment once trained. In this work, we propose the soft Gradient Boosting Machine (sGBM) by wiring multiple differentiable base learners together, by injecting both local and global objectives inspired from gradient boosting, all base learners can then be jointly optimized with linear speed-up. When using differentiable soft decision trees as base learner, such device can be regarded as an alternative version of the (hard) gradient boosting decision trees with extra benefits. Experimental results showed that, sGBM enjoys much higher time efficiency with better accuracy, given the same base learner in both on-line and off-line settings. arXiv:2006.04059v1 [cs.LG] 7 Jun 2020 1. Introduction Gradient Boosting Machine (GBM) [Fri01] has proven to be one successful function approximator and has been widely used in a variety of areas [BL07, CC11]. The basic idea is to train a series of base learners that minimize some predefined differentiable loss function in a sequential fashion.
    [Show full text]
  • Xgboost: a Scalable Tree Boosting System
    XGBoost: A Scalable Tree Boosting System Tianqi Chen Carlos Guestrin University of Washington University of Washington [email protected] [email protected] ABSTRACT problems. Besides being used as a stand-alone predictor, it Tree boosting is a highly effective and widely used machine is also incorporated into real-world production pipelines for learning method. In this paper, we describe a scalable end- ad click through rate prediction [15]. Finally, it is the de- to-end tree boosting system called XGBoost, which is used facto choice of ensemble method and is used in challenges widely by data scientists to achieve state-of-the-art results such as the Netflix prize [3]. on many machine learning challenges. We propose a novel In this paper, we describe XGBoost, a scalable machine learning system for tree boosting. The system is available sparsity-aware algorithm for sparse data and weighted quan- 2 tile sketch for approximate tree learning. More importantly, as an open source package . The impact of the system has we provide insights on cache access patterns, data compres- been widely recognized in a number of machine learning and sion and sharding to build a scalable tree boosting system. data mining challenges. Take the challenges hosted by the machine learning competition site Kaggle for example. A- By combining these insights, XGBoost scales beyond billions 3 of examples using far fewer resources than existing systems. mong the 29 challenge winning solutions published at Kag- gle's blog during 2015, 17 solutions used XGBoost. Among these solutions, eight solely used XGBoost to train the mod- Keywords el, while most others combined XGBoost with neural net- Large-scale Machine Learning s in ensembles.
    [Show full text]
  • Downloaded from the Cancer Imaging Archive (TCIA)1 [13]
    UC Irvine UC Irvine Electronic Theses and Dissertations Title Deep Learning for Automated Medical Image Analysis Permalink https://escholarship.org/uc/item/78w60726 Author Zhu, Wentao Publication Date 2019 License https://creativecommons.org/licenses/by/4.0/ 4.0 Peer reviewed|Thesis/dissertation eScholarship.org Powered by the California Digital Library University of California UNIVERSITY OF CALIFORNIA, IRVINE Deep Learning for Automated Medical Image Analysis DISSERTATION submitted in partial satisfaction of the requirements for the degree of DOCTOR OF PHILOSOPHY in Computer Science by Wentao Zhu Dissertation Committee: Professor Xiaohui Xie, Chair Professor Charless C. Fowlkes Professor Sameer Singh 2019 c 2019 Wentao Zhu DEDICATION To my wife, little son and my parents ii TABLE OF CONTENTS Page LIST OF FIGURES vi LIST OF TABLES x LIST OF ALGORITHMS xi ACKNOWLEDGMENTS xii CURRICULUM VITAE xiii ABSTRACT OF THE DISSERTATION xvi 1 Introduction 1 1.1 Dissertation Outline and Contributions . 3 2 Adversarial Deep Structured Nets for Mass Segmentation from Mammograms 7 2.1 Introduction . 7 2.2 FCN-CRF Network . 9 2.3 Adversarial FCN-CRF Nets . 10 2.4 Experiments . 12 2.5 Conclusion . 16 3 Deep Multi-instance Networks with Sparse Label Assignment for Whole Mammogram Classification 17 3.1 Introduction . 17 3.2 Deep MIL for Whole Mammogram Mass Classification . 19 3.2.1 Max Pooling-Based Multi-Instance Learning . 20 3.2.2 Label Assignment-Based Multi-Instance Learning . 22 3.2.3 Sparse Multi-Instance Learning . 23 3.3 Experiments . 24 3.4 Conclusion . 27 iii 4 DeepLung: Deep 3D Dual Path Nets for Automated Pulmonary Nodule Detection and Classification 29 4.1 Introduction .
    [Show full text]
  • Gradient Boosting Meets Graph Neural Networks
    Published as a conference paper at ICLR 2021 BOOST THEN CONVOLVE:GRADIENT BOOSTING MEETS GRAPH NEURAL NETWORKS Sergei Ivanov Liudmila Prokhorenkova Criteo AI Lab; Skoltech Yandex; HSE University; MIPT Paris, France Moscow, Russia [email protected] [email protected] ABSTRACT Graph neural networks (GNNs) are powerful models that have been successful in various graph representation learning tasks. Whereas gradient boosted decision trees (GBDT) often outperform other machine learning methods when faced with heterogeneous tabular data. But what approach should be used for graphs with tabular node features? Previous GNN models have mostly focused on networks with homogeneous sparse features and, as we show, are suboptimal in the heterogeneous setting. In this work, we propose a novel architecture that trains GBDT and GNN jointly to get the best of both worlds: the GBDT model deals with heterogeneous features, while GNN accounts for the graph structure. Our model benefits from end- to-end optimization by allowing new trees to fit the gradient updates of GNN. With an extensive experimental comparison to the leading GBDT and GNN models, we demonstrate a significant increase in performance on a variety of graphs with tabular features. The code is available: https://github.com/nd7141/bgnn. 1 INTRODUCTION Graph neural networks (GNNs) have shown great success in learning on graph-structured data with various applications in molecular design (Stokes et al., 2020), computer vision (Casas et al., 2019), combinatorial optimization (Mazyavkina et al., 2020), and recommender systems (Sun et al., 2020). The main driving force for progress is the existence of canonical GNN architecture that efficiently encodes the original input data into expressive representations, thereby achieving high-quality results on new datasets and tasks.
    [Show full text]
  • Enhanced Prediction of Hot Spots at Protein-Protein Interfaces Using Extreme Gradient Boosting
    www.nature.com/scientificreports OPEN Enhanced Prediction of Hot Spots at Protein-Protein Interfaces Using Extreme Gradient Boosting Received: 5 June 2018 Hao Wang, Chuyao Liu & Lei Deng Accepted: 10 September 2018 Identifcation of hot spots, a small portion of protein-protein interface residues that contribute the Published: xx xx xxxx majority of the binding free energy, can provide crucial information for understanding the function of proteins and studying their interactions. Based on our previous method (PredHS), we propose a new computational approach, PredHS2, that can further improve the accuracy of predicting hot spots at protein-protein interfaces. Firstly we build a new training dataset of 313 alanine-mutated interface residues extracted from 34 protein complexes. Then we generate a wide variety of 600 sequence, structure, exposure and energy features, together with Euclidean and Voronoi neighborhood properties. To remove redundant and irrelevant information, we select a set of 26 optimal features utilizing a two-step feature selection method, which consist of a minimum Redundancy Maximum Relevance (mRMR) procedure and a sequential forward selection process. Based on the selected 26 features, we use Extreme Gradient Boosting (XGBoost) to build our prediction model. Performance of our PredHS2 approach outperforms other machine learning algorithms and other state-of-the-art hot spot prediction methods on the training dataset and the independent test set (BID) respectively. Several novel features, such as solvent exposure characteristics, second structure features and disorder scores, are found to be more efective in discriminating hot spots. Moreover, the update of the training dataset and the new feature selection and classifcation algorithms play a vital role in improving the prediction quality.
    [Show full text]
  • Development and Evaluation of the Combined Machine Learning Models for the Prediction of Dam Inflow
    water Article Development and Evaluation of the Combined Machine Learning Models for the Prediction of Dam Inflow Jiyeong Hong 1 , Seoro Lee 1, Joo Hyun Bae 2 , Jimin Lee 1, Woon Ji Park 1, Dongjun Lee 1, Jonggun Kim 1 and Kyoung Jae Lim 1,* 1 Department of Regional Infrastructure Engineering, Kangwon National University, Chuncheon-si 24341, Korea; [email protected] (J.H.); [email protected] (S.L.); [email protected] (J.L.); [email protected] (W.J.P.); [email protected] (D.L.); [email protected] (J.K.) 2 Korea Water Environment Research Institute, Chuncheon-si 24408, Korea; [email protected] * Correspondence: [email protected] Received: 3 September 2020; Accepted: 15 October 2020; Published: 20 October 2020 Abstract: Predicting dam inflow is necessary for effective water management. This study created machine learning algorithms to predict the amount of inflow into the Soyang River Dam in South Korea, using weather and dam inflow data for 40 years. A total of six algorithms were used, as follows: decision tree (DT), multilayer perceptron (MLP), random forest (RF), gradient boosting (GB), recurrent neural network–long short-term memory (RNN–LSTM), and convolutional neural network–LSTM (CNN–LSTM). Among these models, the multilayer perceptron model showed the best results in predicting dam inflow, with the Nash–Sutcliffe efficiency (NSE) value of 0.812, root mean squared errors (RMSE) of 77.218 m3/s, mean absolute error (MAE) of 29.034 m3/s, correlation coefficient (R) of 0.924, and determination coefficient (R2) of 0.817. However, when the amount of dam inflow is below 100 m3/s, the ensemble models (random forest and gradient boosting models) performed better than MLP for the prediction of dam inflow.
    [Show full text]
  • On Interpretability of Artificial Neural Networks
    On Interpretability of Artificial Neural Networks: A Survey Feng-Lei Fan, Student Member, IEEE, Jinjun Xiong, Senior Member, IEEE, Mengzhou Li, Student Member, IEEE, and Ge Wang, Fellow, IEEE Abstract— Deep learning as represented by the artificial deep learning, interpretability is needed to hold a deep learning neural networks (DNNs) has achieved great success in many system accountable [57]. If a model builder can explain why a important areas that deal with text, images, videos, graphs, and model makes a particular decision under certain conditions, so on. However, the black-box nature of DNNs has become one of users would know whether such a model contributes to an the primary obstacles for their wide acceptance in mission- adverse event or not. It is then possible to establish standards critical applications such as medical diagnosis and therapy. Due to the huge potential of deep learning, interpreting neural and protocols to use the deep learning system optimally. networks has recently attracted much research attention. In this The lack of interpretability has become a main barrier of deep paper, based on our comprehensive taxonomy, we systematically learning in its wide acceptance in mission-critical applications. review recent studies in understanding the mechanism of neural For example, regulations were proposed by European Union in networks, describe applications of interpretability especially in medicine, and discuss future directions of interpretability 2016 that individuals affected by algorithms have the right to research, such as in relation to fuzzy logic and brain science. obtain an explanation [61]. Despite great research efforts made Index Terms—Deep learning, neural networks, on interpretability of deep learning and availability of several interpretability, survey.
    [Show full text]
  • An Intuitive Exploration of Artificial Intelligence
    An Intuitive Exploration of Artificial Intelligence Simant Dube An Intuitive Exploration of Artificial Intelligence Theory and Applications of Deep Learning Simant Dube Technology and Innovation Office Varian Medical Systems A Siemens Healthineers Company Palo Alto, CA, USA ISBN 978-3-030-68623-9 ISBN 978-3-030-68624-6 (eBook) https://doi.org/10.1007/978-3-030-68624-6 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
    [Show full text]
  • Comparison of Different U-Net Models for Building Extraction from High- Resolution Aerial Imagery Firat ERDEM, Ugur AVDAN
    ISSN: 2148-9173 Vol: 7 Issue:3 Dec 2020 International Journal of Environment and Geoinformatics (IJEGEO) is an international, multidisciplinary, peer reviewed, open access journal. Comparison of Different U-Net Models for Building Extraction from High- Resolution Aerial Imagery Firat ERDEM, Ugur AVDAN Chief in Editor Prof. Dr. Cem Gazioğlu Co-Editors Prof. Dr. Dursun Zafer Şeker, Prof. Dr. Şinasi Kaya, Prof. Dr. Ayşegül Tanık and Assist. Prof. Dr. Volkan Demir Guest Editor Dr. Nedim Onur AYKUT Editorial Committee (December 2020) Assos. Prof. Dr. Abdullah Aksu (TR), Assit. Prof. Dr. Uğur Algancı (TR), Prof. Dr. Bedri Alpar (TR), Prof. Dr. Levent Bat (TR), Prof. Dr. Paul Bates (UK), İrşad Bayırhan (TR), Prof. Dr. Bülent Bayram (TR), Prof. Dr. Luis M. Botana (ES), Prof. Dr. Nuray Çağlar (TR), Prof. Dr. Sukanta Dash (IN), Dr. Soofia T. Elias (UK), Prof. Dr. A. Evren Erginal (TR), Assoc. Prof. Dr. Cüneyt Erenoğlu (TR), Dr. Dieter Fritsch (DE), Prof. Dr. Çiğdem Göksel (TR), Prof.Dr. Lena Halounova (CZ), Prof. Dr. Manik Kalubarme (IN), Dr. Hakan Kaya (TR), Assist. Prof. Dr. Serkan Kükrer (TR), Assoc. Prof. Dr. Maged Marghany (MY), Prof. Dr. Michael Meadows (ZA), Prof. Dr. Nebiye Musaoğlu (TR), Prof. Dr. Masafumi Nakagawa (JP), Prof. Dr. Hasan Özdemir (TR), Prof. Dr. Chryssy Potsiou (GR), Prof. Dr. Erol Sarı (TR), Prof. Dr. Maria Paradiso (IT), Prof. Dr. Petros Patias (GR), Prof. Dr. Elif Sertel (TR), Prof. Dr. Nüket Sivri (TR), Prof. Dr. Füsun Balık Şanlı (TR), Prof. Dr. Uğur Şanlı (TR), Duygu Ülker (TR), Prof. Dr. Seyfettin Taş (TR), Assoc. Prof. Dr. Ömer Suat Taşkın (US), Assist.
    [Show full text]
  • Imitation Learning in Relational Domains: a Functional-Gradient Boosting Approach
    Imitation Learning in Relational Domains: A Functional-Gradient Boosting Approach Sriraam Natarajan and Saket Joshi+ and Prasad Tadepalli+ and Kristian Kersting# and Jude Shavlik∗ Wake Forest University School of Medicine, USA + Oregon State University, USA # Fraunhofer IAIS, Germany, ∗ University of Wisconsin-Madison, USA Abstract follow a meaningful trajectory. Nevertheless, techniques used from supervised learning have been successful for imitation Imitation learning refers to the problem of learn- learning [Ratliff et al., 2006]. We follow this tradition and ing how to behave by observing a teacher in ac- investigate the use of supervised learning methods to learn tion. We consider imitation learning in relational behavioral policies. domains, in which there is a varying number of ob- jects and relations among them. In prior work, sim- Inverse reinforcement learning is a popular approach to ple relational policies are learned by viewing imi- imitation learning studied under the name of apprentice- tation learning as supervised learning of a function ship learning [Abbeel and Ng, 2004; Ratliff et al., 2006; from states to actions. For propositional worlds, Neu and Szepesvari, 2007; Syed and Schapire, 2007]. Here functional gradient methods have been proved to one would assume that the observations are generated by a be beneficial. They are simpler to implement than near-optimal policy of a Markov Decision Process (MDP) most existing methods, more efficient, more natu- with an unknown reward function. The approaches usually rally satisfy common constraints on the cost func- take the form of learning the reward function of the agent and tion, and better represent our prior beliefs about solving the corresponding MDP.
    [Show full text]
  • Episodic Exploration for Deep Deterministic Policies for Starcraft Micromanagement
    Published as a conference paper at ICLR 2017 EPISODIC EXPLORATION FOR DEEP DETERMINISTIC POLICIES FOR STARCRAFT MICROMANAGEMENT Nicolas Usunier∗, Gabriel Synnaeve∗, Zeming Lin, Soumith Chintala Facebook AI Research {usunier,gab,zlin,soumith}@fb.com ABSTRACT We consider scenarios from the real-time strategy game StarCraft as benchmarks for reinforcement learning algorithms. We focus on micromanagement, that is, the short-term, low-level control of team members during a battle. We propose several scenarios that are challenging for reinforcement learning algorithms because the state- action space is very large, and there is no obvious feature representation for the value functions. We describe our approach to tackle the micromanagement scenarios with deep neural network controllers from raw state features given by the game engine. We also present a heuristic reinforcement learning algorithm which combines direct exploration in the policy space and backpropagation. This algorithm collects traces for learning using deterministic policies, which appears much more efficient than, e.g., -greedy exploration. Experiments show that this algorithm allows to successfully learn non-trivial strategies for scenarios with armies of up to 15 agents, where both Q-learning and REINFORCE struggle. 1 INTRODUCTION StarCraft1 is a real-time strategy (RTS) game in which each player must build an army and control individual units to destroy the opponent’s army. As of today, StarCraft is considered one of the most difficult games for computers, and the best bots only reach the level of high amateur human players (Churchill, 2015). The main difficulty comes from the need to control a large number of units in partially observable environment, with very large state and action spaces: for example, in a typical game, there are at least 101685 possible states whereas the game of Go has about 10170 states.
    [Show full text]
  • Scene Classification Via a Gradient Boosting Random Convolutional
    This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 1 Scene Classification via a Gradient Boosting Random Convolutional Network Framework Fan Zhang, Bo Du, Senior Member, IEEE, and Liangpei Zhang, Senior Member, IEEE Abstract—Due to the recent advances in satellite sensors, a large from these VHR images effectively and efficiently has drawn amount of high-resolution remote sensing images is now being great interest in the remote sensing field [1]. However, due to obtained each day. How to automatically recognize and analyze the complex composition and the large number of land-cover scenes from these satellite images effectively and efficiently has become a big challenge in the remote sensing field. Recently, a types, it is still a difficult task. lot of work in scene classification has been proposed, focusing on There have been many attemptsover the yearsto recognize and deep neural networks, which learn hierarchical internal feature analyze scenes from VHR images effectively and efficiently. representations from image data sets and produce state-of-the-art The majority of the early approaches, such as the bag-of-visual- performance. However, most methods, including the traditional words (BoVW) based scene classification method [2], simply shallow methods and deep neural networks, only concentrate on training a single model. Meanwhile, neural network ensembles count the occurrences of the low-level features [e.g., color histo- have proved to be a powerful and practical tool for a number of gram, local binary pattern, and scale-invariant feature transform different predictive tasks.
    [Show full text]